diff --git a/CHANGELOG.md b/CHANGELOG.md
index bf743fa22..84c0017de 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -5,6 +5,7 @@
 ### New Features
 
 ### Bug fixes
+- Fixed a bug where sampled nested hyperparameters (e.g. ``policy_kwargs``) overwrote the whole dictionary during hyperparameter optimization, dropping user-specified keys like ``features_extractor_class`` (@musaaero)
 - Set `learning_starts` to 500 for `parking-v0` in TQC hyperparameters to ensure the first episode is finished
 - Fixed "too many open files" issue when doing hyperparameter optimization: the trial environments are now closed in a `finally` block and `gc_after_trial=True` is passed to Optuna to free them between trials (also needs latest SB3 master version to fully work)
 
diff --git a/rl_zoo3/exp_manager.py b/rl_zoo3/exp_manager.py
index 770a56ba8..b228b00f1 100644
--- a/rl_zoo3/exp_manager.py
+++ b/rl_zoo3/exp_manager.py
@@ -1,4 +1,5 @@
 import argparse
+import copy
 import importlib
 import os
 import pickle as pkl
@@ -60,7 +61,30 @@
     get_wrapper_class,
 )
 
+def deep_update(original: dict, update: dict) -> dict:
+    """
+    Recursively merge ``update`` into ``original`` and return the merged dict.
+
+    Unlike ``dict.update()``, nested dictionaries are merged key-by-key instead of
+    being replaced wholesale, so sampled nested hyperparameters (e.g. ``policy_kwargs``)
+    no longer drop user-specified sibling keys (see issue #431).
+
+    The inputs are not mutated: ``original`` is deep-copied first, which also prevents
+    sampled values from leaking into ``self._hyperparams`` across optimization trials
+    (the previous code used a shallow ``dict.copy()`` that shared nested dicts).
+    """
+    merged = copy.deepcopy(original)
+    _deep_update(merged, update)
+    return merged
+
 
+def _deep_update(target: dict, update: dict) -> None:
+    for key, value in update.items():
+        if isinstance(value, dict) and isinstance(target.get(key), dict):
+            _deep_update(target[key], value)
+        else:
+            target[key] = value
+            
 class ExperimentManager:
     """
     Experiment manager: read the hyperparameters,
@@ -812,7 +836,7 @@ def objective(self, trial: optuna.Trial) -> float:
         # Pass n_actions to initialize DDPG/TD3 noise sampler
         # Sample candidate hyperparameters
         sampled_hyperparams = HYPERPARAMS_SAMPLER[self.algo](trial, self.n_actions, n_envs, additional_args)
-        kwargs.update(sampled_hyperparams)
+        kwargs = deep_update(kwargs, sampled_hyperparams)
 
         env = self.create_envs(n_envs, no_log=True)
 
diff --git a/tests/test_hyperparams_opt.py b/tests/test_hyperparams_opt.py
index f6d5d6e38..89a0f970d 100644
--- a/tests/test_hyperparams_opt.py
+++ b/tests/test_hyperparams_opt.py
@@ -6,6 +6,7 @@
 import optuna
 import pytest
 from optuna.trial import TrialState
+from rl_zoo3.exp_manager import deep_update
 
 
 def _assert_eq(left, right):
@@ -131,3 +132,44 @@ def test_multiple_workers(tmp_path):
 
     for worker in workers:
         assert worker.returncode == 0, "STDOUT:\n{}\nSTDERR:\n{}\n".format(*worker.communicate())
+
+
+
+def test_deep_update_merges_nested_dicts():
+    # Regression test for issue #431: sampled nested hyperparameters must merge
+    # into user-specified nested dicts instead of replacing them wholesale
+    base = {
+        "learning_rate": 3e-4,
+        "policy_kwargs": {
+            "features_extractor_class": "CustomFeatureExtractor",
+            "features_extractor_kwargs": {"features_dim": 10},
+        },
+    }
+    sampled = {
+        "policy_kwargs": {
+            "net_arch": {"pi": [64], "vf": [64]},
+            "activation_fn": "relu",
+        }
+    }
+
+    merged = deep_update(base, sampled)
+
+    # Sampled keys are applied
+    assert merged["policy_kwargs"]["net_arch"] == {"pi": [64], "vf": [64]}
+    assert merged["policy_kwargs"]["activation_fn"] == "relu"
+    # Sibling keys are NOT clobbered (the bug: dict.update() dropped them)
+    assert merged["policy_kwargs"]["features_extractor_class"] == "CustomFeatureExtractor"
+    assert merged["policy_kwargs"]["features_extractor_kwargs"]["features_dim"] == 10
+    # Flat keys behave like before
+    assert merged["learning_rate"] == 3e-4
+    # Inputs are not mutated (the old shallow copy leaked across trials)
+    assert "net_arch" not in base["policy_kwargs"]
+    assert "activation_fn" not in base["policy_kwargs"]
+
+
+def test_deep_update_replaces_non_dict_values():
+    # Merging only happens dict-into-dict; anything else replaces, like dict.update()
+    assert deep_update({"a": {"b": 1}}, {"a": 5}) == {"a": 5}
+    assert deep_update({"a": 1}, {"a": {"b": 2}}) == {"a": {"b": 2}}
+    assert deep_update({}, {"a": 1}) == {"a": 1}
+    assert deep_update({"a": 1}, {}) == {"a": 1}